CDNet: Complementary Depth Network for RGB-D Salient Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Current RGB-D salient object detection (SOD) methods rely on depth data as supplementary information to RGB data.
• Existing RGB-D SOD datasets often contain low-quality depth maps, leading to inaccurate predictions in trained networks.

Purpose of the Study:

• To develop a novel network, the Complementary Depth Network (CDNet), for enhanced RGB-D SOD.
• To address the challenges posed by low-quality depth maps in RGB-D SOD tasks.
• To improve the accuracy and robustness of salient object detection using complementary depth information.

Main Methods:

• Proposed CDNet utilizes saliency-informative depth maps as training targets, leveraging RGB features to estimate improved depth maps.
• Introduced a dynamic fusion scheme with adaptive weights to integrate features from original and estimated depth maps.
• Designed a two-stage cross-modal feature fusion strategy to effectively combine depth and RGB features.

Main Results:

• CDNet demonstrated superior performance compared to state-of-the-art RGB-D SOD methods across seven benchmark datasets.
• The proposed methods for depth map enhancement and feature fusion significantly improved detection accuracy.
• Experimental results validate the effectiveness of CDNet in exploiting depth information for salient object detection.

Conclusions:

• CDNet offers a robust solution for RGB-D SOD by effectively mitigating the impact of low-quality depth data.
• The network's novel feature fusion and depth estimation techniques contribute to state-of-the-art performance.
• The public availability of the code facilitates further research and development in RGB-D SOD.